Byzantine-Robust and Privacy-Preserving Federated Learning With Irregular Participants
Yinuo Chen, Wuzheng Tan, Yijian Zhong, Yulin Kang, Anjia Yang, Jian Weng · IEEE Internet of Things Journal · 2024
Federated learning, as a form of distributed learning, aims to protect the local data while utilizing distributed data to train a global model. However, federated learning still faces challenges related to privacy leakage in Internet of Things (IoT). Researches indicate that the server can infer private information from the local gradients. Additionally, malicious participants may upload poisoning local models, which contaminate the global model and cause a decline in accuracy. Furthermore, irregular participants with low-quality data in the real world can also impact the performance of the global model. Simultaneously addressing these three issues poses a significant challenge. This is because privacy protection strategies in FL are designed to prevent access to the local gradients to avoid information leakage. However, strategies with Byzantine robustness and defense against irregular participants typically require access to the local gradients to calculate the reliability of each participant. Therefore, we use secret sharing as the underlying technology to propose a 3PC privacy-preserving federated learning framework BPFL that can resist Byzantine attacks and irregular participants. Compared with the previous schemes, our scheme can not only protect data privacy but also minimize the negative impact of malicious or irregular participants on the global model. We implemented BPFL and compared it with Mkrum and PPFL. Experimental results indicate that our approach maintains high performance when facing malicious attackers and irregular participants.